We previously reported an elastomeric, optical foam sensor that can sense different types of deformation1 . The elastomeric foam is embedded with optical fibers that shine light into the foam while simultaneously transmitting scattered light exiting the foam. We applied machine learning techniques to the optical fiber data to form a prediction model that predicts whether the foam is being twisted or bent (classification), as well as the magnitude and direction of the deformation (regression). The best classification model had 100% accuracy on new data points, and the best regression models had a mean absolute error of 0.06 degrees on new data points. This kind of proprioceptive ability could give soft robots much more information about their physical state and therefore improve our ability to control them; however, prediction error increases with time due to drift in the optical fiber outputs. This paper presents an attempt to address this drift. We applied a technique based on work presented by Di Carlo et. al2 . This unsupervised technique uses the evolutionary optimization process “covariance matrix adaptation evolution strategy” (CMA-ES) to compute a correction factor that can be applied to unobserved, drifted data points. The best solutions reduced classification error by 49% and regression mean absolute error by 36%.